Predicting of early miscarriage based on machine learning: a retrospective cohort study in China

Abstract

Background

Early miscarriage (EM) remains a significant reproductive health challenge, with current serum hormone biomarkers (β-hCG, progesterone, estradiol) showing inconsistent predictive value across studies. To address these limitations, artificial intelligence (AI) and machine learning (ML) techniques are increasingly applied to integrate complex biomarker data and improve the early prediction of EM.

Methods

This retrospective cohort study analyzed 17,926 singleton pregnancies (1,918 EM vs. 16,008 live births) from Chongqing (2017–2023), China, to establish optimal hormonal predictors and timing for EM risk assessment. Serum hormone concentrations were longitudinally analyzed across distinct gestational weeks to characterize temporal variations. To assess the predictive efficacy of individual biomarkers and composite models, we conducted receiver operating characteristic (ROC) curve analyses. The optimal predictive biomarker and the earliest gestational time point for EM detection were determined through comparative evaluation of area under the ROC curve (AUC) values derived from univariate and multivariate logistic regression models. Finally, a systematic machine learning approach employing ten distinct algorithms was implemented to screen and validate the most robust prediction model.

Results

Longitudinal analysis revealed distinct gestational patterns: β-hCG and estradiol showed weekly increases through 9 weeks before stabilizing, while progesterone demonstrated a biphasic pattern (decline until 7 weeks, subsequent rise). EM cases exhibited significantly lower hormone levels across all gestational weeks. Receiver operating characteristic (ROC) analysis was conducted to assess the predictive performance of both univariate and multivariate models, which revealed a progressive increase in accuracy with advancing gestational age (AUC range: 0.67–0.99). Notably, week 4 progesterone alone showed predictive capacity (AUC = 0.67, cut-off = 17.75 pg/ml). Machine learning evaluation of 10 algorithms identified an artificial neural network (ANN) incorporating β-hCG, progesterone, estradiol, and maternal age as optimal, achieving superior prediction accuracy at week 4 (training AUC = 0.82; validation AUC = 0.81). This multivariate model enables EM risk stratification earlier than conventional hormonal assessments.

Conclusions

Our findings resolve key controversies regarding predictive thresholds and timing while demonstrating the clinical potential of machine learning integration for early pregnancy risk assessment. The established week 4 prediction window offers critical opportunities for timely intervention and patient counseling.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12884-026-09765-z.

Background

Miscarriage, defined as the intrauterine pregnancy loss before 24 weeks gestation, occurs in about 15.3% around the world [1]. Over 90% of miscarriages occur within the first trimester of gestation, which is known as early miscarriage (EM) [2]. Multiple risk factors have been reported to contribute to EM, such as female age, the number of previous pregnant losses, air pollution, smoking, and alcohol, which are always difficult to analyze and investigate [1]. EM can not only increase the risk of successive miscarriages: the risk of miscarriage after three consecutive miscarriages may reach 45%, but also compound the risk of severe ‘miscarriage syndrome’ such as cardiovascular disease, preterm birth, preeclampsia or other obstetrical disorders, causing economic and mental burdens on patients [2]. Therefore, timely and effective prediction for EM is urgent needed.

Serological indicators are reported to be highly sensitive and specific in predicting early pregnancy outcomes, with serum β-hCG, progesterone (P), and estradiol (E2) being the most commonly monitored hormones in the clinic, which are typically lower in women experiencing miscarriage [3, 4]. However, serum hormones vary significantly among individuals, and the detection results are affected by many factors [5]. Meanwhile, due to the differences in patient recruitment criteria, study cohort size, and inclusion gestational age, the meaning, earliest time point and threshold value of serum hormones in predicting early pregnancy outcomes are still controversial [6]. Furthermore, the levels of serum β-hCG, P, and E2 vary each week of pregnancy, which has often been overlooked in previous researches [7, 8]. At present, there is still a lack of large-sample cohort studies to predict pregnancy outcomes by stratification of gestational age. Recently, machine learning (ML) has emerged as powerful tools in obstetrics, capable of analyzing complex, high-dimensional datasets to improve predictive accuracy for adverse outcomes like miscarriage, preterm birth and gestational diabetes [9–11]. These techniques are particularly suited to address the limitations of traditional statistical methods by identifying subtle, non-linear patterns across gestational ages and integrating multiple biomarkers to build robust prognostic models [12].

Here, all singleton pregnant women who had their β-hCG, P, and E2 levels measured at the Chongqing Medical University Affiliated Children’s and Women’s Hospital during the gestational weeks 4^+ 0 to 12^+ 6 between January 1, 2017, and December 31, 2023, were enrolled in our study and divided into the EM and live birth (LB) groups based on pregnancy outcomes. Based on the stratification of one gestational week, we investigated the dynamic changes of serum β-hCG, P, and E2 levels in the first trimester, compared serum hormone levels between the EM and LB groups, evaluated the efficacies of the three serum hormones alone or in combination on the prediction of EM, and determined the earliest gestational week in EM prediction. Then, we selected ‘the earliest gestational week’ patients as the study population, which was randomly divided into the training set and a verification set at a ratio of 7:3. Finally, ten ML algorithms were performed to determine the best predictive model. The detailed research based on a large cohort and stratification of gestational age ensures the reliability of our results, which will greatly contribute to a more scientific utilization of serum hormones for predicting EM and lay the foundation for early treatment of EM.

Methods

Participants

We used electronic medical records of the Women and Children’s Hospital of Chongqing Medical University to collect clinical data of singleton pregnancies who had the laboratory results for β-hCG, P, and E2 levels at gestational weeks 4^+ 0 to 12^+ 6 from Jan 1, 2017 to Dec 31, 2023. Pregnant women carrying a single fetus were diagnosed at 5 weeks by ultrasound scan. Then they were divided into EM and LB groups based on whether they experienced miscarriage within 12 weeks of pregnancy and live birth at delivery. Women with unknown pregnancy outcomes, multiple pregnancies, molar pregnancies, biochemical pregnancies, and ectopic pregnancies were excluded. All hormonal assays (β-hCG, E2, and P) were performed in a single central laboratory using the same automated electrochemiluminescence immunoassay analyzer (ARCHITECT i2000 SR, Abbott Laboratories, USA) and identical reagent kits throughout the study period, ensuring analytical consistency. The inter-assay precision was rigorously monitored using internal quality control procedures. The inter-assay coefficients of variation (CV%) for all analytes were consistently maintained below 8.33%, in accordance with the recommended precision criteria: one-third of the allowable total error (25%) set by the National Center for Clinical Laboratories of China. Finally, the data of 16,008 LB and 1,918 EM were retrospectively studied. The research flow chart is shown in Fig. 1.

Fig. 1.

A flow chart shows the analyses and model-making process for the study

Ethics approval and consent to participate

This study has been approved by the Institutional Review Board (IRB) of Women and Children’s Hospital of Chongqing Medical University and recorded at the China National Medical research registration and filing information system (www.yxyj.org.cn/www.medicalresearch.org.cn, Record number: MR-50-24-014832). All procedures were conducted in accordance with the corresponding guidelines.

For the consent to participate, due to the retrospective nature of the study and the use of anonymized data, the IRB of Women and Children’s Hospital of Chongqing Medical University waived the need of obtaining informed consent.

Study design

We first conducted multiple linear regression to investigate the relationships between maternal age, gestational age, pregnancy outcome, and the levels of serum β-hCG, P, and E2, and compared the differences weekly in serum analytes and maternal age between the two groups. Considering that maternal age may also affect serum analyte levels, we compared the levels of serum markers between the two groups weekly again after stratified random sampling with the observed age distribution of EM as a reference. Later, the receiver operating characteristic (ROC) curve analysis was performed by gestational age to investigate the prediction performance of single or multi-variable based on the data of all participants. Then, the prediction efficiencies of single and multi-indicators were compared to identify the earliest gestational age and the best combination of variables in predicting EM. Finally, data on the ‘earliest gestational age’ was randomly divided into the training set and verification set with a ratio of 7 to 3. Ten machine learning algorithms, namely Adaptive Boosting (AdaBoost), Artificial Neural Network (ANN), Decision Tree (DT), Random Forest (RF), Light Gradient Boosting Machine (LGBM), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Navie Bayesian (NB), and K-Nearest Neighbors (KNN) were introduced to develop the best prediction model for EM using the selected variables above.

Statistical analysis

The Kolmogorov-Smirnov test was employed to assess the normality of all variables (maternal age and the levels of serum β-hCG, P, and E2). As a result, the four variables were non-normally distributed, so we used the Wilcoxon rank-sum test to perform subsequent differential analysis. To match the maternal age in different pregnancy outcome groups, we performed stratified random sampling in the LB group with the observed maternal age distribution of EM as a reference. Later, the Wilcoxon rank-sum test was performed again to determine the difference in serum analytes between the two groups. For ROC analysis of a single indicator, the raw value was directly used as the test variable. For the compound model, we used multiple logistic regression to estimate a prognostic score (predicted probabilities), which was then employed as the test variable in the ROC analysis. Test performance characteristics, including area under curve (AUC), sensitivity, and specificity, were assessed by R package pROC v1.18.5 [13]. The best statistical cut-off values were obtained based on the Youden index, which maximizes the sum of sensitivity and specificity. The 95% confidence intervals (CIs) of AUCs, sensitivities, and specificities were calculated using 2000 stratified bootstrap replicates via the ci.sp function of the pROC package. Then, the P value was calculated using the pt function of R v4.3.2 to evaluate the significance of the AUC from random (AUC = 0.5).

To further explore the predictive performance of multi-variable model for early miscarriage, we employed ten machine learning (ML) algorithms. All models were constructed using Python v3.11 as following steps: (1) Participants were randomly divided into a training set and validation set at a ratio of 7:3; (2) A grid search for hyperparameter optimization of all models using 5-fold cross-validation were then performed on the training set (except for NB algorithm, because it does not require hyperparameter optimization); (3) The predictive performance of the best predictive model was verified on the validation set using ROC analysis. P values of < 0.05 were considered statistically significant. Statistical analysis was performed with R v4.3.2.

Results

Weekly serum levels of β-hCG, P, and E2 during the first trimester

Through retrospective analysis of the clinical data of pregnant women who were in the first trimester in the past six years (2017–2023), a total of 17,925 pregnant women, 16,008 for live births (LB) and 1,918 for early miscarriage (EM) were finally included in the analysis. Consistent with the earlier report [1], we found the incidence of EM is 10.7% (1,918 / 17,925). Multiple linear regression was then performed to evaluate the relationship of these serum analytes (dependent variables) with gestational age, maternal age, and pregnancy outcome (independent variables). As Table 1 shows, three serum analyte levels were all significantly associated with gestational age, maternal age, and pregnancy outcome. And we found that during the first trimester, β-hCG and E2 levels increased significantly with gestational age at 4–9 weeks, then stabilized (Table S1, P > 0.05). Meanwhile, P levels decreased slightly from weeks 4 to 7 with gestational age, followed by a significant increase from weeks 7 to 9, then stabilized (Table S1, P > 0.05). Moreover, levels of the three serum analytes were significantly decreased in the EM group during the first trimester, except for β-hCG and E2 at four weeks of gestation (Fig. 2 and Table S2). However, further analysis revealed an obvious difference in maternal age between the two groups at most gestational weeks (Table S3). Since maternal age was significantly associated with all three serum analyte levels (Table 1), the differences in β-hCG, P, and E2 levels between the two groups should be compared again after eliminating the interference of maternal age. So, we used a stratified random sample to match the maternal age distribution of the two groups in each gestational week, and the ages of each group after matching were shown in Table S4. Consistently, there were significant differences in β-hCG, P, and E2 levels between the two groups after maternal age matching (Table S5 and Figure S1). Together, our results indicate significant differences in β-hCG, P, and E2 levels between the EM and LB groups during the first trimester.

Table 1.

Result of multiple linear regression

Dependent variable	Independent variable	β	Std. Error	t value	P value
β-hCG	Gestational age	21873.31	235.46	92.90	0.00E + 00
	Group*	67149.61	1331.76	50.42	0.00E + 00
	Maternal age	576.51	109.83	5.25	1.55E-07
E2	Gestational age	96.13	4.05	23.71	1.74E-115
	Group*	450.33	22.53	19.99	2.79E-84
	Maternal age	3.95	1.96	2.01	4.45E-02
P	Gestational age	-0.10	0.03	-2.91	3.62E-03
	Group*	7.62	0.19	40.96	0.00E + 00
	Maternal age	0.14	0.02	9.14	6.99E-20

* Refers to results grouped by pregnancy outcome at 12 weeks, i.e., LB group and EM group

LB live birth, EM early miscarriage

Fig. 2.

Weekly comparison of serum analytes between EM and LB groups. a β-hCG levels; b E2 levels; c P levels. EM, early miscarriage. LB, live birth

ROC curve analysis of serum β‑hCG, P, and E2 levels

To clarify the ability of serum analytes to predict EM, we calculated the AUC of the three serum analytes separately at different gestational ages, and the ROC curve was displayed in Fig. 3a-c. The AUC of β-hCG, P, and E2 increased with gestational age and showed significant predictive values at 5 to 12 weeks of gestation (P < 0.05, Table 2). However, at four weeks of pregnancy, only the P showed predictive performance for EM with an AUC of 0.67 (Compared with AUC = 0.5, P < 0.05, Table 2). Figure 3d showed that the prediction efficiency of β-hCG and E2 was generally higher than that of P during the first trimester. But at four weeks of gestation, the AUC value of P was greater than β-hCG and E2. E2 was more recommended for EM prediction at 5–6 weeks of pregnancy, and the AUC value of β-hCG was significantly higher than the other two serum analytes at 7–10 weeks. Moreover, at six weeks, the AUC of E2 was 0.81, with 95% CI 0.77–0.89, indicating excellent predictive accuracy. The AUC of β-hCG reached 0.90 at eight weeks, with 95% CI 0.88–0.91, indicating prominent predictive accuracy. Together, our results suggest that all three serum analytes played essential roles in EM prediction, and it is necessary to combine them to predict EM in the first trimester of pregnancy.

Fig. 3.

Predicted probability of β-hCG, P, and E2 in prediction of EM. EM, early miscarriage. a Predicted probability of β-hCG alone, b Predicted probability of P alone, c Predicted probability of E2 alone, d AUC values of β-hCG, P, and E2 with increasing gestational ages

Table 2.

Optimal cut-off points and diagnostic performance of serum analytes for distinguishing LB from EM at different gestational ages

Markers	Gestational age, wk	No.	Cut off	Sensitivity% (95% CI)	Specificity% (95% CI)	AUC (95% CI)	P value
β-hCG (mIU/ml)	4	772	8902.6	91.94 (85.48–98.39)	17.75 (14.93–20.7)	0.54 (0.47–0.61)	2.43E-01
	5	1913	4880.3	53.74 (45.58–61.22)	67.89 (65.68–69.99)	0.62 (0.57–0.66)	1.48E-06
	6	2614	27198.1	74.7 (69.17–79.84)	68.11 (66.24–69.93)	0.76 (0.73–0.79)	2.32E-67
	7	3811	59583.3	75.33 (71.56–79.56)	76.47 (75.04–77.98)	0.83 (0.81–0.85)	8.41E-224
	8	3666	74293.45	81.14 (77.68–84.58)	86.61 (85.38–87.8)	0.9 (0.88–0.91)	0.00E + 00
	9	2558	76831.8	84.74 (80.69–88.47)	88.91 (87.57–90.25)	0.91 (0.89–0.93)	0.00E + 00
	10	1330	73005.15	79.26 (71.85–85.93)	90.04 (88.28–91.72)	0.9 (0.86–0.93)	1.62E-98
	11	485	73224.65	90 (76.67–100)	88.57 (85.71–91.43)	0.93 (0.87–0.99)	4.57E-46
	12	124	57567.1	83.33 (58.33–100)	90.18 (83.93–95.54)	0.92 (0.85–0.99)	3.62E-35
P (ng/ml)	4	722	17.75	41.38 (29.31–55.17)	88.55 (86.14–90.96)	0.67 (0.59–0.75)	4.25E-05
	5	1884	20.25	52.74 (44.52–60.96)	66.69 (64.44–68.87)	0.61 (0.56–0.66)	3.61E-05
	6	2601	18.45	58.06 (52.01–64.52)	74.12 (72.42–75.9)	0.72 (0.68–0.75)	1.56E-32
	7	3840	17.25	61.52 (57.27-66)	79.34 (77.98–80.67)	0.75 (0.72–0.78)	1.46E-72
	8	3767	17.85	71.75 (67.42–75.46)	79.19 (77.79–80.56)	0.81 (0.78–0.83)	2.56E-137
	9	2628	17.55	74.38 (69.06–79.06)	82.37 (80.76–83.97)	0.85 (0.83–0.88)	1.69E-153
	10	1328	16.05	67.65 (59.56-75)	90.77 (89.09–92.45)	0.85 (0.81–0.89)	3.76E-67
	11	480	16.25	77.42 (61.29–90.32)	91.09 (88.42–93.55)	0.85 (0.74–0.95)	4.55E-10
	12	120	16.75	100 (100–100)	86.11 (79.63–92.59)	0.96 (0.93-1)	3.11E-143
E2 (pg/ml)	4	249	229.5	63.64 (36.36–90.91)	74.79 (69.75–80.25)	0.63 (0.43–0.84)	2.04E-01
	5	584	355	86.11 (74.93–97.22)	58.39 (54.38–62.41)	0.76 (0.68–0.84)	1.40E-10
	6	600	354	80.88 (70.59–89.71)	72.74 (68.98–76.32)	0.83 (0.77–0.89)	3.72E-29
	7	568	425.5	75 (66.3–83.7)	76.68 (72.69–80.46)	0.82 (0.77–0.87)	3.46E-39
	8	437	556.5	90.2 (84.31–95.1)	67.46 (62.69–72.24)	0.87 (0.83–0.91)	2.45E-74
	9	278	419	80.36 (69.64–91.07)	81.98 (76.58–86.94)	0.87 (0.81–0.93)	1.85E-36
	10	116	443.5	72.73 (54.55–90.91)	84.04 (76.6-91.49)	0.84 (0.75–0.94)	4.54E-12
	11	34	307	100 (100–100)	89.29 (78.48–100)	0.94 (0.86-1)	1.84E-26

ROC curve analysis of multiple markers

The multivariate logistic regression was used to evaluate the predictive power of β-hCG, P, E2, and maternal age in combination. The optimal cut-off values and predictive performances of multi-indicators for EM prediction were displayed by gestational week in Table S6. Overall, the predictive efficacy increased with gestational age. The model containing maternal age and β-hCG showed relatively poor predictive power while containing β-hCG, P, E2, and maternal age obtained the best predictive performance (Fig. 4). Meanwhile, introducing the variable of maternal age into the logistic regression model had no significant impact on the predictive efficacy, but in early pregnancy, especially in the fourth week of pregnancy, adding maternal age to our model could increase the AUC by 0.04. In detail, the AUC combined with β-hCG and E2 alone reached 0.85 (95% CI 0.80–0.90) at gestational week 6, and with the gestational age arising to week 7, the AUC value combining all variables increased to 0.87 (95% CI 0.83–0.91). At gestational week 8, the predictive model based on all variables reached an AUC of 0.90 (95% CI 0.86–0.93), which increased to 0.93 (95% CI 0.87–0.98) at ten weeks of gestation. Additionally, at gestational week 11, all models containing β-hCG and P could achieve the highest AUC value of 0.99 (Fig. 4 and Table S6).

Fig. 4.

The heatmap of AUC values with multi-indicators

Comparison of ROC curves between single and multiple combined variables

The best AUC values for predicting EM by single and multiple indicators were displayed in Fig. 5 at each gestational week. Overall, the AUC values of multiple variables in the first trimester were higher than those of the single serum analyte, except for β-hCG with an AUC of 0.91 (95% CI 0.89–0.93) at the 9th week of gestation. In detail, at gestational week 4, P showed a best AUC of 0.67 (95% CI 0.59–0.75) among all serum analytes, while combining all variables provided the highest AUC with 0.75 (95% CI 0.60–0.90). The maximum AUC of gestational age five was only 0.02 higher than that of gestational age four. In comparison, the maximum AUC at gestational age six was higher than that at five gestational weeks, showing an excellent predictive efficacy (AUC = 0.85, 95% CI: 0.81–0.91). Moreover, the best AUC value at eight gestational age showed prominent predictive efficacy (0.90, 95% CI: 0.86–0.94). Together, our results showed that combining multi-indicators can advance the earliest time point for EM prediction to four weeks of pregnancy.

Fig. 5.

Comparison of best AUC values between the single and multiple indicators at each gestational age

Construction and verification of the optimal multi-variable model for EM prediction based on ten machine learning algorithms at four weeks gestation

To further explore the predictive performance of multi-variable combinations for EM prediction, we obtained 110 classification models using ten machine learning (ML) algorithms (Fig. 6). In summary, using cross-validation, we found that the ANN model, with an AUC of 0.82 (95% CI: 0.79 − 0.86) in the training set, showed an AUC of 0.81 (95% CI: 0.75 − 0.87) in the validation set (Fig. 6a and Table S7), which was obviously superior to other ML models (Fig. 6b and Table S7).

Fig. 6.

The predictive efficiencies of 110 classification models based on the fourth gestational week data using ten machine learning (ML) algorithms. a AUC values of the training set, b AUC values of the validation set. The bar chart on the far right represents the maximum AUC obtained by combining different algorithms based on the same metric. Adaptive Boosting (AdaBoost), Artificial Neural Network (ANN), Decision Tree (DT), Random Forest (RF), Light Gradient Boosting Machine (LGBM), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Navie Bayesian (NB), K-Nearest Neighbors (KNN)

Comparison to previous studies

We searched studies about past twenty years from PubMed with the keywords “(((predict[Title/Abstract]) OR (diagnosis[Title/Abstract])) AND ((((HCG[Title/Abstract]) OR (progesterone[Title/Abstract])) OR (estradiol[Title/Abstract])) AND (((pregnancy outcome[Title/Abstract]) OR (pregnancy loss[Title/Abstract]) OR (abortion[Title/Abstract])) OR (miscarriage[Title/Abstract]))))”, and found 525 records. The researches most relevant to our study are shown in Table 3. As we can see, among these studies [7, 8, 14–17], our study cohort is the largest in the field of EM prediction and has the most detailed grouping (Table 3). What’s more, as shown in Fig. 7, our study demonstrated overall higher predictive performance compared to other studies. However, in comparison to the researches by Deng et. al [4] (gestational weeks 5–6) and Li et.al [6] (gestational weeks 7), which had smaller samples than our study, the predictive efficacies of our study are lower. Finally, as far as we know, our study is the first to demonstrate that the ANN model containing serum hormones and maternal age at 4 weeks gestation have excellent predictive power (AUC = 0.80) for EM.

Table 3.

Gestational ages, patient numbers, predictive performances and serum indicators from previous studies in early miscarriage prediction

Study	Title	Gestational age		No. (miscarriage vs. control)	Indicator	AUC
Kassie J. Bollig et al. (2023) [18]	Validation of a multiple marker test for early pregnancy outcome prediction	5–10	127 vs. 65		Activin A	0.58 [0.49, 0.66]
					PAPP-A	0.61 [0.52, 0.69]
					Glycodelin	0.53 [0.43, 0.63]
					ADAM12	0.59 [0.51, 0.68]
					P	0.84 [0.79, 0.90]
					hCG	0.81 [0.74, 0.89]
Yinglin Liu et al. (2016) [19]	Predictive value of serum β-hCG for early pregnancy outcomes among women with recurrent spontaneous abortion	5–13	358 vs. 882		hCG-peak	0.948 [0.930, 0.965]
Jesper Friis Petersen et al. (2020) [20]	Soluble urokinase plasminogen activator receptor (suPAR) as a biomarker of early pregnancy location and viability compared with hCG, progesterone and estradiol	< 6	18 vs. 85		hCG	0.70 [0.60, 0.80]
					E2	0.75 [0.64, 0.85]
					P	0.84 [0.75, 0.92]
		> 6	18 vs. 79		hCG	0.96 [0.93, 0.99]
					E2	0.92 [0.88, 0.96]
					P	0.88 [0.83, 0.93]
Yi Gong et al. (2024) [21]	Can single progesterone concentration predict miscarriage in early pregnant women with threatened miscarriage: a systematic review and meta-analysis	4–12	3106 (total, from 15 studies)		P	0.85 [0.81, 0.88]
Yang Li et al. (2021) [22]	Significance of dynamically monitoring serum estrogen and β-human chorionic gonadotropin in early pregnancy assessment	5	36 vs. 118		E2	0.69 [0.59, 0.79]
		6	41 vs. 128		E2	0.82 [0.75, 0.90]
		7	41 vs. 137		E2	0.90 [0.83, 0.96]
		8	24 vs. 94		E2	0.88 [0.80, 0.96]
Wenhui Deng et al. (2022) [23]	Prediction of miscarriage in first trimester by serum estradiol, progesterone and β-human chorionic gonadotropin within 9 weeks of gestation	5–6	61 vs. 85		E2 + P + hCG	0.869 [0.759, 0.980]
					E2 + P	0.871 [0.793, 0.950]
					E2 + hCG	0.795 [0.641, 0.948]
					P + hCG	0.865 [0.760, 0.968]
					E2	0.866 [0.793, 0.938]
					P	0.766 [0.672, 0.861]
					hCG	0.661
		7–9	46 vs. 65		E2 + P + hCG	0.768 [0.670, 0.866]
					E2 + P	0.739 [0.651, 0.827]
					E2 + hCG	0.770 [0.672, 0.869]
					P + hCG	0.667 [0.561, 0.772]
					E2	0.709 [0.616, 0.801]
					P	0.529
					hCG	0.396
Ilias Katsikis et al. (2006) [24]	Receiver operator characteristics and diagnostic value of progesterone and CA-125 in the prediction of ectopic and abortive intrauterine gestations	3–10	20 vs. 20		P	0.859 [0.768, 0.951]
					CA-125	0.788 [0.639, 0.957]

Fig. 7.

Comparison of the predictive performance between previous studies and our study. Logistic Regression Model, LR. Artificial Neural Network, ANN

Discussion

This retrospective cohort study systematically evaluated the changes in serum β-hCG, progesterone (P), and estradiol (E2) levels with advancing gestational age and developed an artificial neural network (ANN) model that integrates these serum biomarkers along with maternal age, which may serve as an early risk stratification tool at 4 weeks of gestation. Meanwhile, based on the data of total 17,925 pregnancies over six years, we found that the incidence of EM is 10.7% (1,918 / 17,925), the serum β-hCG and E2 levels significantly increased with gestational age at 4–9 weeks, stabilized at 9–12 weeks, and the serum β-hCG, P, and E2 levels in EM group were significantly lower than those in the live birth (LB) group during the first trimester, which were consistent with previous studies [6, 17, 25]. However, there were also some unexpected discoveries. Our result showed that P levels decreased slightly with gestational age at 4–7 weeks, increased significantly at 7–9 weeks, and then stabilized. However, Deng et al. observed a different result. By comparing the serum P levels between the groups of 5–6 weeks gestation and 7–9 weeks gestation, they believed that P levels increased with gestational age [17]. The difference in these results may be due to grouping patients with multiple gestational weeks together. In fact, Neumann et al. pointed out that the production of significant P typically begins around the 7th gestational week, and P increase is non-linear and depends on chorionicity and zygosity [25]. What’s more, we found that at the four weeks of gestation, only P showed statistically significant predictive value, although the clinical significance of an AUC of 0.67 may be modest. In fact, high-risk patients could be considered for closer monitoring or early interventions, such as P supplementation, which has shown efficacy in preventing miscarriage in selected populations [26–30]. To some extent, our study here also provides the basis for P supplementation in early pregnancy. However, any clinical application should be mindful that our model was developed using specific laboratory methods: serum levels of β-hCG, P, and E2 were measured using the Abbott i2000 SR automated electrochemiluminescence immunoassay analyzer. Institutions using different assays would need to validate appropriate thresholds accordingly.

Previous studies showed that maternal age was one of the critical factors affecting serum hormone levels and pregnancy outcome [31, 32]. Here, we also found that maternal age was obviously higher in the EM group than in the LB group. Our study found that regardless of maternal age matching, the differences in serum hormone levels between the EM and the LB groups were consistent across each gestational week. This could be attributed to the fact that in our study, the median age of most pregnant women was between 28 and 31 years old, potentially masking the potential impact of a smaller subset of mothers aged below 20 or above 40 on the serum hormone levels and the risk of pregnancy outcome, as this subset was reported to be more prone to miscarriage [32–34]. It is worth noting that maternal age is still included as a variable in the final machine learning (ML) model for enrolling all variables in the logistic regression model yielded the highest AUC, especially at four gestational weeks, and that recording maternal ages has no adverse impact on patients’ health or economic status.

There has been a longstanding issue perplexing researchers in pregnancy outcome prediction, namely, the cut-off values varied largely among different studies [35–37]. To solve this problem, we investigated the cut-off values of β-hCG, E2, and P respectively at each gestational age from 4 to 12 weeks. We found that the cut-off values varied with different gestational weeks (Table 2). What’s more, our results also showed that serum P at four weeks gestation, rather than β-hCG or E2, has the potential to predict EM. When using the single indicator, E2 was more recommended for EM prediction at 5–6 weeks of pregnancy, and the AUC value of β-hCG was significantly higher than the other two serum analytes at 7–10 weeks. Together, our findings suggested that patients at different gestational weeks cannot be grouped indiscriminately, which was often overlooked in most previous studies [7, 8, 14–17]. In certain studies where grouping across gestational weeks is necessary, patients should be grouped based on the relationship between hormone levels and gestational weeks, specifically incorporating gestational weeks with similar hormone levels into the same group.

Last but not least, we compared our study with previous researches on the prediction of EM [6–8, 10–13], and found our study has the largest cohort study with the largest number of subgroups, which demonstrating the earliest time point for predicting early miscarriage as the fourth week of pregnancy.

However, this study still remains several limitations. Firstly, this study is unicentric. Although internal validation showed promising performance, the model has not been externally validated in other populations or healthcare settings. Therefore, its generalizability may be limited. Moreover, due to the retrospective nature of the study, several potential confounding factors that may influence hormone levels and pregnancy outcomes. While we controlled for maternal age through stratified sampling, other factors such as luteal phase support, maternal comorbidities (e.g., polycystic ovary syndrome, thyroid disorders), body mass index, and lifestyle factors were not accounted for in our analysis. Thus, the results may not be sufficiently comprehensive. Further considerable, high-quality, and multicentric prospective research is necessary to validate and refine these preliminary findings for improved clinical relevance.

Conclusion

Conclusively, in our study, we were mainly concerned with the changes in the levels of three hormones with increasing gestational age, the relationships between these hormone levels and EM prediction, the earliest time point at which serum indicators are effective for EM prediction, and the best predictive model based on the combination of multiple indicators. Our study demonstrated that the earliest time point in EM prediction was four weeks of gestation, at this time, the ANN model containing β-hCG, P, E2, and maternal age showed excellent predictive efficacy on EM (AUC = 0.81).

Abbreviations

EM

Early miscarriage

ROC

Receiver operating characteristic curve

AUC

Area under receiver-operating characteristics curve

ANN

Artificial neural network

LB

Live birth

IRB

Institutional Review Board

P

Progesterone

E2

Estradiol

CIs

Confidence intervals

ML

Machine learning

Authors’ contributions

QL, JC, and TJC contributed to the conception and design of the study. JLQ taking full responsibility for leading the revision process. JC, HZ, and DCZ were responsible for data collection. QL, TJC, FHL, and SL contributed to the analysis and interpretation of data. QL, JLQ, and TJC drafted the paper. HZ, QL, and CLL revised the paper for important intellectual content. QL and CLL approved the final version of the manuscript to be submitted.

Funding

This work was supported by the Science-Health Joint Medical Scientific Research Project of Chongqing (2024QNXM019), the Youth Project of Science and Technology Research Program of Chongqing Municipal Education Commission (Grant No. KJQN202600444), and Natural Science Foundation of Chongqing Municipality (CN) (CSTB2023NSCQ-BHX0025).

Data availability

The data that support the findings of this study are available from the corresponding author, upon reasonable request.

Declarations

Ethics approval and consent to participate

This study has been approved by the Institutional Review Board (IRB) of Women and Children’s Hospital of Chongqing Medical University and recorded at the China National Medical research registration and filing information system (www.yxyj.org.cn/www.medicalresearch.org.cn, Record number: MR-50-24-014832). All procedures were conducted in accordance with the corresponding guidelines. For the consent to participate, due to the retrospective nature of the study and the use of anonymized data, the IRB of Women and Children’s Hospital of Chongqing Medical University waived the need of obtaining informed consent.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.